Predictive AI with your own data: demand, late payments and no-show appointments

The artificial intelligence that makes the news writes: it drafts emails, summarises documents, answers questions. There is another kind, far less eye-catching, that has been working for years in banks, insurers and large distributors: the kind that predicts from the tables your business software already holds.

Predictive AI learns from your own history (the rows and columns of your ERP, your CRM or your appointment software) and calculates how likely something is to happen: that an invoice will be paid late, that a patient will miss an appointment, how much each customer will order next week. Its job is to rank: it tells you what to look at first. A few hundred or a few thousand cases are enough to try it, and its value lies in the decision it prompts, reviewed by a person.

Two kinds of AI: one that writes and one that decides

Generative AI produces content: text, images, code. Predictive AI works with tabular data, that is, information arranged in rows and columns, like a spreadsheet or the tables in business software. Each row is a case (an invoice, an appointment, an order) and each column a feature (the amount, the customer, the day of the week, the days late on the last payment).

The two complement each other. Predictive AI calculates. Generative AI explains the result in plain language, drafts the email or creates the task for the person who has to act.

There is a wide gap between what this technology can do and what companies actually use. In McKinsey's 2025 survey, 88% of companies say they use AI regularly in at least one area, but fewer than two in five attribute any improvement in operating results to AI. Part of that gap is right here: the history is already in the business software, and nobody turns it into a daily decision.

What it can forecast from the data you already have

These are the most common uses in mid-sized companies, and they all come from data that already exists:

• Demand by customer or by product. How much each customer will order, or how much of each item you will sell next week. It helps you buy better and avoid running out of your fastest movers.
• Late payments and defaults. Which invoices are most likely to be paid late, based on how each customer has paid so far. You call the riskiest first, before the due date.
• Appointments nobody turns up to. In clinics, workshops, advisory firms or training centres, which appointments are most likely to be left empty. You confirm those more firmly or offer the slot to the waiting list.
• Quotes that will be accepted. Which are most likely to close and which are going cold, so follow-up goes where it pays off most.
• Customers who are drifting away. Who has bought less often lately, before they stop buying altogether.
• Stock that expires or runs out. Which items will expire before they are used at the current rate of consumption.

In every case the question is the same: what decision do you make every day or every week, with what data, and what does getting it wrong cost.

How it works, step by step

1. The history is gathered. Past cases with their known outcome: invoices with the actual payment date, appointments with whether the patient came, orders with what was delivered. It comes out of the business software, without changing it.
2. The model learns and is tested against the past. It is trained on one period (say, 2024) and measured on another it has not seen (2025). If it does not beat what you already do by eye, it is not used.
3. Every day it scores what is new. This week's invoices, tomorrow's appointments, the open quotes. Next to each probability it shows which factors explain it.
4. Generative AI explains it and turns it into tasks. "These five invoices have a high risk of being late: two are from customers who took longer over their last three payments." It then drafts the reminder or the task for whoever has to act.
5. A person decides. The prediction organises the work and the decision stays with your team.

How much data you need

Less than people tend to think. Until recently, a predictive model needed a lot of history and weeks of tuning. Tabular foundation models, pre-trained on millions of synthetic datasets, have changed that. The best known, TabPFN, published in Nature in January 2025 by a team led by the University of Freiburg, outperforms earlier methods on tables of up to 10,000 rows and returns results in seconds.

For a mid-sized company, that means two or three years of its own history can be enough for a first test. What you do need is for the outcome to be recorded: if nobody notes whether the patient came or the actual date payment arrived, there is nothing to learn from.

What it does not do

• It gives probabilities. An 80% risk means that out of ten similar cases, eight went badly. The other two turned out fine.
• It learns from the past. If something fundamental changes (a crisis in a sector, a very large new customer, a different price list), the model is slow to notice. Its accuracy has to be measured every month.
• It finds relationships, but does not explain their causes. That Monday orders are delivered later more often does not say why.
• It inherits the biases in the history. If one type of customer was treated worse in the past, the model will learn it. That is why the factors it uses are reviewed.

What the law says

If the prediction affects people, two rules are worth keeping in mind. The GDPR (Article 22) gives people the right not to be subject to decisions based solely on automated processing that produce legal effects or similarly significant effects: with a person reviewing each decision, the model organises the work and does not decide alone. And the EU AI Act classes as high-risk the systems that assess the creditworthiness of natural persons. Forecasting demand, stock or late payments by other companies does not fall into that category. The application dates of the AI Act are in our AI glossary (in Spanish).

The data also does not have to leave your environment: the model can run with the data in the European Union (in Spanish) or on your own servers.

How to start

Pick a single recurring decision that costs money today when it goes wrong: invoices paid late, empty appointment slots or stock that expires. A proof of concept is built with your history and compared with what actually happened before it is used day to day.

At Tecnea, the proof of concept starts at €850, full development ranges from €5,000 to €20,000 and the monthly fee starts at €250 (what each amount includes, in Spanish). If your business software lets its data be read, it does not need replacing: we explain it in does AI integrate with my ERP? (in Spanish). And if you are torn between the AI built into your software and your own development, the criteria are in is the AI built into my software enough?.

Frequently asked questions

Do I need a data team to use predictive AI?

No. Whoever develops the application prepares and maintains the model. Your team sees the result where it already works (a list ranked by risk, an alert, a task) and decides.

How much data do I need?

For a first test, a few hundred or a few thousand cases with their recorded outcome are usually enough, for example two or three years of invoices or appointments. What matters is that the outcome is written down.

Do I have to change my business software?

No. The data is read from the software you already use, and the result goes back to it or to the tool your team works in.

How is it different from a dashboard?

A dashboard shows what has already happened. A predictive model estimates what will happen with each specific case (this invoice, this appointment) and ranks them so you can act earlier.

Can it decide on its own, without a person?

Technically yes, but we do not recommend it when people are affected, and the GDPR limits fully automated decisions with significant effects. The usual approach is for the model to organise the work and a person to decide.

If you have a recurring decision and a history in your business software, tell us what it is and we will tell you whether it can be predicted from your data.

This article is for information. Tecnea develops custom AI applications, predictive ones included, so we have an interest in you trying it. That is why we also explain what it does not do and when the law requires a person behind each decision.

Sources

1. McKinsey & Company, "The state of AI in 2025: Agents, innovation, and transformation" (November 2025): mckinsey.com
2. Hollmann, N. et al., "Accurate predictions on small data with a tabular foundation model", Nature 637, 319-326 (January 2025): nature.com
3. University of Freiburg, "New AI model TabPFN enables faster and more accurate predictions on small tabular data sets" (January 2025): uni-freiburg.de
4. Regulation (EU) 2016/679 (GDPR), Article 22: eur-lex.europa.eu
5. Regulation (EU) 2024/1689 on artificial intelligence, Annex III: eur-lex.europa.eu